Overview
Emdash is an experimental coding agent system that explores parallel execution through git worktrees. Instead of running one agent at a time, Emdash spins up multiple agent instances on separate worktrees, each tackling a different task or approach. This concurrency model promises significant speedups for roadmap execution and multi-file refactoring.
| Attribute | Value |
|---|---|
| Language / Stack | TypeScript / Node.js |
| GitHub | github.com/emdash-io/emdash |
| License | MIT |
| Closest Match | Parallel agent execution + git worktrees |
| Maturity | Experimental |
Key Features
- Git worktree orchestration — Automatically creates and manages multiple worktrees for parallel agent sessions.
- Task decomposition — Breaks large PRs into independent subtasks that can run concurrently.
- Conflict resolution — Detects and merges changes from parallel worktrees using automated three-way merge.
- Resource pooling — Manages compute budget across parallel agents to prevent rate limit exhaustion.
- Result aggregation — Collects outputs from all agents and surfaces the best solution or a combined patch.
Architecture
Emdash uses a coordinator-worker pattern. The coordinator maintains a queue of tasks and spawns worker agents in separate git worktrees. Each worker operates on an isolated branch created from a common base. The coordinator monitors progress via file-system watchers and git hooks. When workers complete, their branches are merged through a preference-weighted algorithm that resolves conflicts using LLM-based diff analysis.
Strengths
- Parallel speedup — Theoretical 3-5x throughput improvement for independent tasks.
- Git-native isolation — Worktrees provide clean separation without Docker overhead.
- Fault tolerance — Failure of one agent does not block others.
- Experimentation-friendly — Easy to try multiple approaches and pick the best result.
Gaps vs Open SWE
- Experimental stability — Merge conflicts are not always resolved correctly; manual intervention sometimes required.
- High compute cost — Running 3-5 agents in parallel multiplies API costs proportionally.
- Not SWE-bench validated — No published benchmark results comparing parallel vs serial execution.
- Limited scope — Works best for refactorings and additive changes; less suitable for complex architectural redesigns.
Relevance to xCoder
Emdash represents our future direction for high-throughput autopilot mode. While our current implementation runs agents serially, we have designed the FlowEngine with parallel phase execution in mind. Emdash's worktree coordination algorithm will inform our M2 roadmap for multi-agent orchestration. We do not plan to adopt Emdash directly, but we are tracking its merge-resolution techniques for our own multi-branch PR automation.
Research priority